Deep Dive

Recursive Learning Methodology

How VIOS agents improve with every decision

The AutoResearch Pattern

The VIOS learning loop is based on the autoresearch pattern: a single-objective optimization loop where agents execute, evaluate their results against a benchmark, adjust parameters, and iterate.

Unlike traditional ML training which requires large datasets and manual intervention, the autoresearch pattern allows agents to continuously improve through operational feedback. Every invoice processed, every contract reviewed, every risk scored becomes a training signal that compounds over time.

The key insight is constraint: each iteration changes one variable, measures the impact, and either keeps the change or reverts it. This mechanical discipline prevents the kind of drift that plagues systems that try to optimize everything at once.

The 4-Stage Cycle

01

Execute

Agents perform their assigned tasks — invoice reconciliation, contract review, risk scoring — using current knowledge and parameters.

02

Score

The Quality Assurance Agent evaluates each output against defined benchmarks: accuracy, completeness, timeliness, and business impact.

03

Learn

The Memory & Learning Agent extracts patterns from scored outputs. Successful patterns are added to the knowledge base; failures are analyzed for root cause.

04

Improve

The Workforce Optimizer Agent adjusts agent parameters, updates decision thresholds, and refines routing logic based on accumulated learning.

ExecuteScoreLearnImproverepeat

Measurable Results

Each iteration targets a specific improvement vector. The table below shows what changed at each step and the resulting accuracy gain.

Ver.
Acc.
What Changed
v1
72%
Initial model deployment. Baseline accuracy established across 847 vendor records.
v2
78%
Added rate card validation rules. Invoice matching accuracy improved by 6pp after processing 2,400 invoices.
v3
84%
Incorporated SLA breach pattern detection. Contract review false positives reduced by 34%.
v4
89%
Cross-agent learning activated. Spend Analytics findings now inform Contract Review decisions.
v5
92%
Vendor risk scoring model updated with Q1 financial data. Risk prediction lead time extended to 47 days.
v6
94%
Full recursive loop stabilized. 12,847 patterns captured. Self-healing auto-approval rate reached 78%.
Δ
+22pp
Total improvement across 6 learning cycles

Architecture

The recursive learning engine operates as a background process across the entire VIOS agent workforce. Every agent output feeds into the evaluation pipeline, and every evaluation informs the next execution cycle.

Cycle Duration

~2 hours

Each learning cycle takes approximately 2 hours

Patterns Captured

12,847

Total patterns captured since deployment

Pattern Categories

5

Rate validation, SLA monitoring, Risk signals, Contract clauses, Spend anomalies

Evaluation Benchmark

VMO Std v4.2

Industry-standard vendor management quality framework

References

Inspired by the autoresearch pattern (Karpathy, 2024) — single-file optimization loops for autonomous improvement. The pattern demonstrates that constrained, single-objective iteration cycles can achieve continuous improvement without manual retraining pipelines.

github.com/karpathy/autoresearch

The VIOS implementation extends the original pattern to multi-agent orchestration: instead of a single file optimizing a single metric, an entire workforce of specialized agents collectively optimizes vendor management outcomes across spend, risk, contract quality, and partnership value simultaneously.

Back to Agents

VIOS RECURSIVE LEARNING